MBZUAI’s IFM Releases PAN 2, an Open 29B World Model for Physical AI

What changed

MBZUAI’s Institute of Foundation Models (IFM) launched PAN 2, an open-source 29 billion parameter world model aimed at physical AI applications. It combines a large language model (LLM) planner with a 29B diffusion transformer to simulate physical environments. PAN 2 can generate video predictions of up to 20 seconds showing possible outcomes of actions, making it a significant step beyond static or limited-horizon world models. The entire system is released under the Apache 2.0 license, making it accessible for experimentation and integration.

Why builders should care

World models that link language-driven planning with rich physical simulation have been tough to develop at scale. PAN 2 provides a versatile, open framework for research and product teams focused on embodied AI, robotics, or interactive environments. Simulating action consequences visually over multiple seconds lets agents plan with foresight, reducing costly trial-and-error in the real world. The model’s architecture balances scale and complexity, making it a practical candidate for labs or startups aiming to build physical reasoning agents without starting from scratch.

The practical takeaway

Combining an LLM with a video diffusion backbone lets PAN 2 interpret high-level goals as detailed sequences of physical state changes, visible through generated video. This can accelerate robotics simulations, automated testing of robot plans, or virtual prototyping of physical tasks. Its open license means operators can tailor the model or retrain it for specific domains, potentially lowering the barrier to entry for physical AI innovation. However, running a 29B parameter model at video scale requires significant GPU resources, so cost and engineering complexity remain considerations.

What to watch next

Adoption of PAN 2 will depend on how well it integrates with existing robotics frameworks and real-time control systems. Watch for third-party tools or libraries that wrap its output into actionable motion commands. Expect MBZUAI or the open-source community to release further benchmarks comparing PAN 2’s planning accuracy and sample efficiency against proprietary models. Also monitor developments in hardware optimizations that might make this class of large visual world models more affordable for commercial use.

AI Quick Briefs Editorial Desk